Fast-adaptive early-stage remaining useful life prediction of lithium-ion batteries with meta-learning

An accurate remaining useful life (RUL) prediction of lithium-ion batteries is the cornerstone for precautionary usage planning. Although versatile, data-driven models rely heavily on the availability of large amounts of high-quality data, which is particularly critical for battery RUL prediction tasks. In this work, we propose a meta-learning framework for fast-adaptive early-stage RUL prediction of lithium-ion batteries using a gated recurrent unit (GRU)-based encoder–decoder model with model-agnostic meta-learning (MAML), exploiting only a limited number of battery degradation profiles to train a meta-model with a good initialization that can quickly adapt to the aging patterns of new cells and carry out robust recursive multi-step trajectory extrapolation into the future. Test results show that the proposed framework is able to thrive on the scarce degradation data and rapidly adapt to new cells, outperforming the baseline and transfer learning models in most usage scenarios, even and especially for long-horizon recursive prediction covering the deep-degradation regime with an MAE of 1.156%. In-depth analyses are presented regarding the influence of the lengths of the prediction horizon and the historical data available for training. We believe this work could shed a new light on cell-specific RUL prediction in the research field as well as industrial applications. • A meta-learning framework is proposed for fast-adaptive early-stage RUL prediction. • A GRU-based encoder–decoder model is developed for robust recursive SOH forecasting. • Insights are presented on the influence of the prediction horizon and history length.